mirror of
https://github.com/onestardao/WFGY.git
synced 2026-04-30 12:39:55 +00:00
92 lines
3.5 KiB
Markdown
92 lines
3.5 KiB
Markdown
# 📒 Long‑Context Stress Problem Map
|
||
|
||
Mega‑prompts—>100 k tokens, entire book dumps, OCR‑noisy PDFs—overwhelm ordinary LLM pipelines.
|
||
WFGY keeps reasoning stable with adaptive ΔS, chunk‑mapping, and sliding Tree windows.
|
||
|
||
---
|
||
|
||
## 🤔 Typical Long‑Context Crashes
|
||
|
||
| Stressor | What Standard Systems Do |
|
||
|----------|--------------------------|
|
||
| 100 k+ tokens | Memory wipe or truncated output |
|
||
| Mixed domains | Topic bleed, incoherent jumps |
|
||
| Duplicate sections | Infinite loops / “as mentioned above” spam |
|
||
| OCR noise | Hallucination or garbage sentences |
|
||
|
||
---
|
||
|
||
## 🛡️ WFGY Countermeasures
|
||
|
||
| Stressor | WFGY Module | Remedy | Status |
|
||
|----------|-------------|--------|--------|
|
||
| 100 k+ tokens | **Chunk‑Mapper** + Sliding Tree | Splits doc into ΔS‑balanced chunks, streams into window | 🛠 Beta |
|
||
| Mixed domains | Per‑domain ΔS fork | Separate Tree branch per domain; no bleed | ✅ |
|
||
| Duplicate sections | **BBMC** dedupe scan | Detects near‑identical residue, collapses | ✅ |
|
||
| PDF OCR noise | BBMC noise filter | Drops >80 % low‑entropy lines | ✅ |
|
||
|
||
---
|
||
|
||
## ✍️ Demo — 150 k‑Token PDF Dump
|
||
|
||
```txt
|
||
1️⃣ Start
|
||
> Start
|
||
|
||
2️⃣ Upload huge PDF text
|
||
> [paste or stream]
|
||
|
||
WFGY process:
|
||
• Chunk‑Mapper splits into 8 k‑token slices
|
||
• For each slice: ΔS calc → Tree node → sliding window
|
||
• Duplicate residue removed (413 sections merged)
|
||
• OCR noise filtered (ΔS noise gate at 0.8)
|
||
• Final summary or Q&A runs with stable context
|
||
````
|
||
|
||
---
|
||
|
||
## 🛠 Module Cheat‑Sheet
|
||
|
||
| Module | Role |
|
||
| ----------------------- | ---------------------------------- |
|
||
| **Chunk‑Mapper** | Adaptive split by semantic tension |
|
||
| **Sliding Tree Window** | Keeps only relevant slices active |
|
||
| **ΔS Metric** | Guides chunk size & window hop |
|
||
| **BBMC** | Dedupe + noise filter |
|
||
| **BBPF** | Forks domain branches if needed |
|
||
|
||
---
|
||
|
||
## 📊 Implementation Status
|
||
|
||
| Feature | State |
|
||
| ------------------- | --------------------- |
|
||
| Chunk‑Mapper | 🛠 Beta (public soon) |
|
||
| Sliding Tree window | ✅ Stable |
|
||
| Cross‑domain fork | ✅ Stable |
|
||
| OCR noise filter | ✅ Stable |
|
||
| GUI chunk viewer | 🔜 Planned |
|
||
|
||
---
|
||
|
||
## 📝 Tips & Limits
|
||
|
||
* For >150 k tokens, set `chunk_max = 6k` for faster pass.
|
||
* Use `tree pause` to inspect each domain branch before auto‑merge.
|
||
* Share monster PDFs in **Discussions**—they stress‑test Chunk‑Mapper.
|
||
|
||
---
|
||
|
||
### 🔗 Quick‑Start Downloads (60 sec)
|
||
|
||
| Tool | Link | 3‑Step Setup |
|
||
| -------------------------- | --------------------------------------------------- | ---------------------------------------------------------------------------------------- |
|
||
| **WFGY 1.0 PDF** | [Engine Paper](https://zenodo.org/records/15630969) | 1️⃣ Download · 2️⃣ Upload to LLM · 3️⃣ Ask “Summarize using WFGY + \<doc>” |
|
||
| **TXT OS (plain‑text OS)** | [TXTOS.txt](https://zenodo.org/records/15788557) | 1️⃣ Download · 2️⃣ Paste into any LLM chat · 3️⃣ Type “hello world” — OS boots instantly |
|
||
|
||
---
|
||
|
||
> Survived a 100k‑token dump? ⭐ the repo so Chunk‑Mapper hits v1.0 faster.
|
||
> ↩︎ [Back to Problem Index](../README.md)
|
||
|